GUIDE: A preliminary study on NLP-based personalized support for type 1 diabetes management

In collaboration with Enhance-d

Managing Type 1 Diabetes (T1D) requires constant attention to glucose levels, especially during physical activities. Recent advancements in natural language processing (NLP) offer new possibilities for providing personalized recommendations based on data from patients’ activities. However, challenges remain in ensuring that generated advice is accurate, contextually appropriate, and relevant.

In our work, we explored the feasibility of using NLP models to generate personalized, actionable advice for T1D patients based on their physical activity. We developed a set of prompts to guide the models and evaluated their ability to produce recommendations that are both medically sound and tailored to the individual’s condition. We tested several open-source models and refined our approach iteratively.

Our results (Mitrović et al., 2025) show that models such as Mistral perform well in generating meaningful insights for managing T1D. In contrast, other models, such as JSL-MedPhi2, performed poorly. This study demonstrates the feasibility of the task and highlights the importance of careful prompt design and model selection in health-related NLP applications, while suggesting that further work is needed to address safety and generalizability concerns.